freebsd-nq/sys/netinet/tcp_lro.c

977 lines
22 KiB
C
Raw Normal View History

/*-
* Copyright (c) 2007, Myricom Inc.
* Copyright (c) 2008, Intel Corporation.
* Copyright (c) 2012 The FreeBSD Foundation
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
* Copyright (c) 2016 Mellanox Technologies.
* All rights reserved.
*
* Portions of this software were developed by Bjoern Zeeb
* under sponsorship from the FreeBSD Foundation.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
2008-06-11 22:12:50 +00:00
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_inet.h"
#include "opt_inet6.h"
2008-06-11 22:12:50 +00:00
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/malloc.h>
#include <sys/mbuf.h>
2008-06-11 22:12:50 +00:00
#include <sys/socket.h>
#include <sys/sysctl.h>
2008-06-11 22:12:50 +00:00
#include <net/if.h>
#include <net/if_var.h>
2008-06-11 22:12:50 +00:00
#include <net/ethernet.h>
#include <net/vnet.h>
2008-06-11 22:12:50 +00:00
#include <netinet/in_systm.h>
#include <netinet/in.h>
#include <netinet/ip6.h>
2008-06-11 22:12:50 +00:00
#include <netinet/ip.h>
#include <netinet/ip_var.h>
2008-06-11 22:12:50 +00:00
#include <netinet/tcp.h>
#include <netinet/tcp_lro.h>
#include <netinet/tcp_var.h>
2008-06-11 22:12:50 +00:00
#include <netinet6/ip6_var.h>
2008-06-11 22:12:50 +00:00
#include <machine/in_cksum.h>
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
2008-06-11 22:12:50 +00:00
#define TCP_LRO_UPDATE_CSUM 1
#ifndef TCP_LRO_UPDATE_CSUM
#define TCP_LRO_INVALID_CSUM 0x0000
#endif
2008-06-11 22:12:50 +00:00
static void tcp_lro_rx_done(struct lro_ctrl *lc);
static int tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m,
uint32_t csum, int use_hash);
SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
"TCP LRO");
static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
"default number of LRO entries");
static __inline void
tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
struct lro_entry *le)
{
LIST_INSERT_HEAD(&lc->lro_active, le, next);
LIST_INSERT_HEAD(bucket, le, hash_next);
}
static __inline void
tcp_lro_active_remove(struct lro_entry *le)
{
LIST_REMOVE(le, next); /* active list */
LIST_REMOVE(le, hash_next); /* hash bucket */
}
2008-06-11 22:12:50 +00:00
int
tcp_lro_init(struct lro_ctrl *lc)
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
{
return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
}
int
tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
unsigned lro_entries, unsigned lro_mbufs)
2008-06-11 22:12:50 +00:00
{
struct lro_entry *le;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
size_t size;
unsigned i, elements;
2008-06-11 22:12:50 +00:00
lc->lro_bad_csum = 0;
lc->lro_queued = 0;
lc->lro_flushed = 0;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
lc->lro_mbuf_count = 0;
lc->lro_mbuf_max = lro_mbufs;
lc->lro_cnt = lro_entries;
lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
lc->ifp = ifp;
LIST_INIT(&lc->lro_free);
LIST_INIT(&lc->lro_active);
2008-06-11 22:12:50 +00:00
/* create hash table to accelerate entry lookup */
if (lro_entries > lro_mbufs)
elements = lro_entries;
else
elements = lro_mbufs;
lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
HASH_NOWAIT);
if (lc->lro_hash == NULL) {
memset(lc, 0, sizeof(*lc));
return (ENOMEM);
}
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* compute size to allocate */
size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
(lro_entries * sizeof(*le));
lc->lro_mbuf_data = (struct lro_mbuf_sort *)
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
malloc(size, M_LRO, M_NOWAIT | M_ZERO);
/* check for out of memory */
if (lc->lro_mbuf_data == NULL) {
free(lc->lro_hash, M_LRO);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
memset(lc, 0, sizeof(*lc));
return (ENOMEM);
}
/* compute offset for LRO entries */
le = (struct lro_entry *)
(lc->lro_mbuf_data + lro_mbufs);
/* setup linked list */
for (i = 0; i != lro_entries; i++)
LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
return (0);
2008-06-11 22:12:50 +00:00
}
void
tcp_lro_free(struct lro_ctrl *lc)
2008-06-11 22:12:50 +00:00
{
struct lro_entry *le;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
unsigned x;
/* reset LRO free list */
LIST_INIT(&lc->lro_free);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* free active mbufs, if any */
while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
tcp_lro_active_remove(le);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
m_freem(le->m_head);
}
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* free hash table */
free(lc->lro_hash, M_LRO);
lc->lro_hash = NULL;
lc->lro_hashsz = 0;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* free mbuf array, if any */
for (x = 0; x != lc->lro_mbuf_count; x++)
m_freem(lc->lro_mbuf_data[x].mb);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
lc->lro_mbuf_count = 0;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* free allocated memory, if any */
free(lc->lro_mbuf_data, M_LRO);
lc->lro_mbuf_data = NULL;
}
2008-06-11 22:12:50 +00:00
#ifdef TCP_LRO_UPDATE_CSUM
static uint16_t
tcp_lro_csum_th(struct tcphdr *th)
{
uint32_t ch;
uint16_t *p, l;
ch = th->th_sum = 0x0000;
l = th->th_off;
p = (uint16_t *)th;
while (l > 0) {
ch += *p;
p++;
ch += *p;
p++;
l--;
}
while (ch > 0xffff)
ch = (ch >> 16) + (ch & 0xffff);
return (ch & 0xffff);
}
static uint16_t
tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
uint16_t tcp_data_len, uint16_t csum)
{
uint32_t c;
uint16_t cs;
c = csum;
/* Remove length from checksum. */
switch (le->eh_type) {
#ifdef INET6
case ETHERTYPE_IPV6:
{
struct ip6_hdr *ip6;
ip6 = (struct ip6_hdr *)l3hdr;
if (le->append_cnt == 0)
cs = ip6->ip6_plen;
else {
uint32_t cx;
cx = ntohs(ip6->ip6_plen);
cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
}
break;
}
#endif
#ifdef INET
case ETHERTYPE_IP:
{
struct ip *ip4;
ip4 = (struct ip *)l3hdr;
if (le->append_cnt == 0)
cs = ip4->ip_len;
else {
cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
IPPROTO_TCP);
cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
htons(cs));
}
break;
}
#endif
default:
cs = 0; /* Keep compiler happy. */
2008-06-11 22:12:50 +00:00
}
cs = ~cs;
c += cs;
/* Remove TCP header csum. */
cs = ~tcp_lro_csum_th(th);
c += cs;
while (c > 0xffff)
c = (c >> 16) + (c & 0xffff);
return (c & 0xffff);
2008-06-11 22:12:50 +00:00
}
#endif
2008-06-11 22:12:50 +00:00
static void
tcp_lro_rx_done(struct lro_ctrl *lc)
{
struct lro_entry *le;
while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
tcp_lro_active_remove(le);
tcp_lro_flush(lc, le);
}
}
void
tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
{
struct lro_entry *le, *le_tmp;
struct timeval tv;
if (LIST_EMPTY(&lc->lro_active))
return;
getmicrotime(&tv);
timevalsub(&tv, timeout);
LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
if (timevalcmp(&tv, &le->mtime, >=)) {
tcp_lro_active_remove(le);
tcp_lro_flush(lc, le);
}
}
}
2008-06-11 22:12:50 +00:00
void
tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
2008-06-11 22:12:50 +00:00
{
if (le->append_cnt > 0) {
struct tcphdr *th;
uint16_t p_len;
p_len = htons(le->p_len);
switch (le->eh_type) {
#ifdef INET6
case ETHERTYPE_IPV6:
{
struct ip6_hdr *ip6;
ip6 = le->le_ip6;
ip6->ip6_plen = p_len;
th = (struct tcphdr *)(ip6 + 1);
le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
CSUM_PSEUDO_HDR;
le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
break;
2008-06-11 22:12:50 +00:00
}
#endif
#ifdef INET
case ETHERTYPE_IP:
{
struct ip *ip4;
#ifdef TCP_LRO_UPDATE_CSUM
uint32_t cl;
uint16_t c;
#endif
ip4 = le->le_ip4;
#ifdef TCP_LRO_UPDATE_CSUM
/* Fix IP header checksum for new length. */
c = ~ip4->ip_sum;
cl = c;
c = ~ip4->ip_len;
cl += c + p_len;
while (cl > 0xffff)
cl = (cl >> 16) + (cl & 0xffff);
c = cl;
ip4->ip_sum = ~c;
#else
ip4->ip_sum = TCP_LRO_INVALID_CSUM;
#endif
ip4->ip_len = p_len;
th = (struct tcphdr *)(ip4 + 1);
le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
le->p_len += ETHER_HDR_LEN;
break;
}
#endif
default:
th = NULL; /* Keep compiler happy. */
}
le->m_head->m_pkthdr.csum_data = 0xffff;
le->m_head->m_pkthdr.len = le->p_len;
/* Incorporate the latest ACK into the TCP header. */
th->th_ack = le->ack_seq;
th->th_win = le->window;
/* Incorporate latest timestamp into the TCP header. */
if (le->timestamp != 0) {
uint32_t *ts_ptr;
ts_ptr = (uint32_t *)(th + 1);
ts_ptr[1] = htonl(le->tsval);
ts_ptr[2] = le->tsecr;
}
#ifdef TCP_LRO_UPDATE_CSUM
/* Update the TCP header checksum. */
le->ulp_csum += p_len;
le->ulp_csum += tcp_lro_csum_th(th);
while (le->ulp_csum > 0xffff)
le->ulp_csum = (le->ulp_csum >> 16) +
(le->ulp_csum & 0xffff);
th->th_sum = (le->ulp_csum & 0xffff);
th->th_sum = ~th->th_sum;
#else
th->th_sum = TCP_LRO_INVALID_CSUM;
#endif
2008-06-11 22:12:50 +00:00
}
le->m_head->m_pkthdr.lro_nsegs = le->append_cnt + 1;
(*lc->ifp->if_input)(lc->ifp, le->m_head);
lc->lro_queued += le->append_cnt + 1;
lc->lro_flushed++;
bzero(le, sizeof(*le));
LIST_INSERT_HEAD(&lc->lro_free, le, next);
2008-06-11 22:12:50 +00:00
}
#ifdef HAVE_INLINE_FLSLL
#define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
#else
static inline uint64_t
tcp_lro_msb_64(uint64_t x)
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
{
x |= (x >> 1);
x |= (x >> 2);
x |= (x >> 4);
x |= (x >> 8);
x |= (x >> 16);
x |= (x >> 32);
return (x & ~(x >> 1));
}
#endif
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/*
* The tcp_lro_sort() routine is comparable to qsort(), except it has
* a worst case complexity limit of O(MIN(N,64)*N), where N is the
* number of elements to sort and 64 is the number of sequence bits
* available. The algorithm is bit-slicing the 64-bit sequence number,
* sorting one bit at a time from the most significant bit until the
* least significant one, skipping the constant bits. This is
* typically called a radix sort.
*/
static void
tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
{
struct lro_mbuf_sort temp;
uint64_t ones;
uint64_t zeros;
uint32_t x;
uint32_t y;
repeat:
/* for small arrays insertion sort is faster */
if (size <= 12) {
for (x = 1; x < size; x++) {
temp = parray[x];
for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
parray[y] = parray[y - 1];
parray[y] = temp;
}
return;
}
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* compute sequence bits which are constant */
ones = 0;
zeros = 0;
for (x = 0; x != size; x++) {
ones |= parray[x].seq;
zeros |= ~parray[x].seq;
}
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* compute bits which are not constant into "ones" */
ones &= zeros;
if (ones == 0)
return;
/* pick the most significant bit which is not constant */
ones = tcp_lro_msb_64(ones);
/*
* Move entries having cleared sequence bits to the beginning
* of the array:
*/
for (x = y = 0; y != size; y++) {
/* skip set bits */
if (parray[y].seq & ones)
continue;
/* swap entries */
temp = parray[x];
parray[x] = parray[y];
parray[y] = temp;
x++;
}
KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
/* sort zeros */
tcp_lro_sort(parray, x);
/* sort ones */
parray += x;
size -= x;
goto repeat;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
}
void
tcp_lro_flush_all(struct lro_ctrl *lc)
{
uint64_t seq;
uint64_t nseq;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
unsigned x;
/* check if no mbufs to flush */
if (lc->lro_mbuf_count == 0)
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
goto done;
/* sort all mbufs according to stream */
tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* input data into LRO engine, stream by stream */
seq = 0;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
for (x = 0; x != lc->lro_mbuf_count; x++) {
struct mbuf *mb;
/* get mbuf */
mb = lc->lro_mbuf_data[x].mb;
/* get sequence number, masking away the packet index */
nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* check for new stream */
if (seq != nseq) {
seq = nseq;
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* flush active streams */
tcp_lro_rx_done(lc);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
}
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* add packet to LRO engine */
if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* input packet to network layer */
(*lc->ifp->if_input)(lc->ifp, mb);
lc->lro_queued++;
lc->lro_flushed++;
}
}
done:
/* flush active streams */
tcp_lro_rx_done(lc);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
lc->lro_mbuf_count = 0;
}
#ifdef INET6
static int
tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
struct tcphdr **th)
2008-06-11 22:12:50 +00:00
{
/* XXX-BZ we should check the flow-label. */
/* XXX-BZ We do not yet support ext. hdrs. */
if (ip6->ip6_nxt != IPPROTO_TCP)
return (TCP_LRO_NOT_SUPPORTED);
/* Find the TCP header. */
*th = (struct tcphdr *)(ip6 + 1);
return (0);
}
#endif
#ifdef INET
static int
tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
struct tcphdr **th)
{
int csum_flags;
uint16_t csum;
if (ip4->ip_p != IPPROTO_TCP)
return (TCP_LRO_NOT_SUPPORTED);
/* Ensure there are no options. */
if ((ip4->ip_hl << 2) != sizeof (*ip4))
return (TCP_LRO_CANNOT);
/* .. and the packet is not fragmented. */
if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
return (TCP_LRO_CANNOT);
/* Legacy IP has a header checksum that needs to be correct. */
csum_flags = m->m_pkthdr.csum_flags;
if (csum_flags & CSUM_IP_CHECKED) {
if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
lc->lro_bad_csum++;
return (TCP_LRO_CANNOT);
}
} else {
csum = in_cksum_hdr(ip4);
if (__predict_false((csum) != 0)) {
lc->lro_bad_csum++;
return (TCP_LRO_CANNOT);
}
2008-06-11 22:12:50 +00:00
}
/* Find the TCP header (we assured there are no IP options). */
*th = (struct tcphdr *)(ip4 + 1);
return (0);
}
#endif
static int
tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, int use_hash)
{
struct lro_entry *le;
struct ether_header *eh;
#ifdef INET6
struct ip6_hdr *ip6 = NULL; /* Keep compiler happy. */
#endif
#ifdef INET
struct ip *ip4 = NULL; /* Keep compiler happy. */
#endif
struct tcphdr *th;
void *l3hdr = NULL; /* Keep compiler happy. */
uint32_t *ts_ptr;
tcp_seq seq;
int error, ip_len, l;
uint16_t eh_type, tcp_data_len;
struct lro_head *bucket;
int force_flush = 0;
/* We expect a contiguous header [eh, ip, tcp]. */
eh = mtod(m, struct ether_header *);
eh_type = ntohs(eh->ether_type);
switch (eh_type) {
#ifdef INET6
case ETHERTYPE_IPV6:
{
CURVNET_SET(lc->ifp->if_vnet);
if (V_ip6_forwarding != 0) {
/* XXX-BZ stats but changing lro_ctrl is a problem. */
CURVNET_RESTORE();
return (TCP_LRO_CANNOT);
}
CURVNET_RESTORE();
l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
if (error != 0)
return (error);
tcp_data_len = ntohs(ip6->ip6_plen);
ip_len = sizeof(*ip6) + tcp_data_len;
break;
}
#endif
#ifdef INET
case ETHERTYPE_IP:
{
CURVNET_SET(lc->ifp->if_vnet);
if (V_ipforwarding != 0) {
/* XXX-BZ stats but changing lro_ctrl is a problem. */
CURVNET_RESTORE();
return (TCP_LRO_CANNOT);
}
CURVNET_RESTORE();
l3hdr = ip4 = (struct ip *)(eh + 1);
error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
if (error != 0)
return (error);
ip_len = ntohs(ip4->ip_len);
tcp_data_len = ip_len - sizeof(*ip4);
break;
}
#endif
/* XXX-BZ what happens in case of VLAN(s)? */
default:
return (TCP_LRO_NOT_SUPPORTED);
2008-06-11 22:12:50 +00:00
}
/*
* If the frame is padded beyond the end of the IP packet, then we must
* trim the extra bytes off.
*/
l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
if (l != 0) {
if (l < 0)
/* Truncated packet. */
return (TCP_LRO_CANNOT);
m_adj(m, -l);
}
/*
* Check TCP header constraints.
*/
/* Ensure no bits set besides ACK or PSH. */
if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
if (th->th_flags & TH_SYN)
return (TCP_LRO_CANNOT);
/*
* Make sure that previously seen segements/ACKs are delivered
* before this segement, e.g. FIN.
*/
force_flush = 1;
}
2008-06-11 22:12:50 +00:00
/* XXX-BZ We lose a ACK|PUSH flag concatenating multiple segments. */
/* XXX-BZ Ideally we'd flush on PUSH? */
/*
* Check for timestamps.
* Since the only option we handle are timestamps, we only have to
* handle the simple case of aligned timestamps.
2008-06-11 22:12:50 +00:00
*/
l = (th->th_off << 2);
tcp_data_len -= l;
l -= sizeof(*th);
ts_ptr = (uint32_t *)(th + 1);
if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
(*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
/*
* Make sure that previously seen segements/ACKs are delivered
* before this segement.
*/
force_flush = 1;
}
/* If the driver did not pass in the checksum, set it now. */
if (csum == 0x0000)
csum = th->th_sum;
seq = ntohl(th->th_seq);
if (!use_hash) {
bucket = &lc->lro_hash[0];
} else if (M_HASHTYPE_ISHASH(m)) {
bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
} else {
uint32_t hash;
switch (eh_type) {
#ifdef INET
case ETHERTYPE_IP:
hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
break;
#endif
#ifdef INET6
case ETHERTYPE_IPV6:
hash = ip6->ip6_src.s6_addr32[0] +
ip6->ip6_dst.s6_addr32[0];
hash += ip6->ip6_src.s6_addr32[1] +
ip6->ip6_dst.s6_addr32[1];
hash += ip6->ip6_src.s6_addr32[2] +
ip6->ip6_dst.s6_addr32[2];
hash += ip6->ip6_src.s6_addr32[3] +
ip6->ip6_dst.s6_addr32[3];
break;
#endif
default:
hash = 0;
break;
}
hash += th->th_sport + th->th_dport;
bucket = &lc->lro_hash[hash % lc->lro_hashsz];
}
/* Try to find a matching previous segment. */
LIST_FOREACH(le, bucket, hash_next) {
if (le->eh_type != eh_type)
continue;
if (le->source_port != th->th_sport ||
le->dest_port != th->th_dport)
continue;
switch (eh_type) {
#ifdef INET6
case ETHERTYPE_IPV6:
if (bcmp(&le->source_ip6, &ip6->ip6_src,
sizeof(struct in6_addr)) != 0 ||
bcmp(&le->dest_ip6, &ip6->ip6_dst,
sizeof(struct in6_addr)) != 0)
continue;
break;
#endif
#ifdef INET
case ETHERTYPE_IP:
if (le->source_ip4 != ip4->ip_src.s_addr ||
le->dest_ip4 != ip4->ip_dst.s_addr)
continue;
break;
#endif
2008-06-11 22:12:50 +00:00
}
if (force_flush) {
/* Timestamps mismatch; this is a FIN, etc */
tcp_lro_active_remove(le);
tcp_lro_flush(lc, le);
return (TCP_LRO_CANNOT);
}
/* Flush now if appending will result in overflow. */
if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
tcp_lro_active_remove(le);
tcp_lro_flush(lc, le);
break;
}
/* Try to append the new segment. */
if (__predict_false(seq != le->next_seq ||
(tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
/* Out of order packet or duplicate ACK. */
tcp_lro_active_remove(le);
tcp_lro_flush(lc, le);
return (TCP_LRO_CANNOT);
}
if (l != 0) {
uint32_t tsval = ntohl(*(ts_ptr + 1));
/* Make sure timestamp values are increasing. */
/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
if (__predict_false(le->tsval > tsval ||
*(ts_ptr + 2) == 0))
return (TCP_LRO_CANNOT);
le->tsval = tsval;
le->tsecr = *(ts_ptr + 2);
}
2008-06-11 22:12:50 +00:00
le->next_seq += tcp_data_len;
le->ack_seq = th->th_ack;
le->window = th->th_win;
le->append_cnt++;
#ifdef TCP_LRO_UPDATE_CSUM
le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
tcp_data_len, ~csum);
#endif
if (tcp_data_len == 0) {
m_freem(m);
/*
* Flush this LRO entry, if this ACK should not
* be further delayed.
*/
if (le->append_cnt >= lc->lro_ackcnt_lim) {
tcp_lro_active_remove(le);
tcp_lro_flush(lc, le);
}
return (0);
2008-06-11 22:12:50 +00:00
}
le->p_len += tcp_data_len;
/*
* Adjust the mbuf so that m_data points to the first byte of
* the ULP payload. Adjust the mbuf to avoid complications and
* append new segment to existing mbuf chain.
*/
m_adj(m, m->m_pkthdr.len - tcp_data_len);
m_demote_pkthdr(m);
le->m_tail->m_next = m;
le->m_tail = m_last(m);
/*
* If a possible next full length packet would cause an
* overflow, pro-actively flush now.
*/
if (le->p_len > (lc->lro_length_lim - lc->ifp->if_mtu)) {
tcp_lro_active_remove(le);
tcp_lro_flush(lc, le);
} else
getmicrotime(&le->mtime);
return (0);
2008-06-11 22:12:50 +00:00
}
if (force_flush) {
/*
* Nothing to flush, but this segment can not be further
* aggregated/delayed.
*/
return (TCP_LRO_CANNOT);
}
/* Try to find an empty slot. */
if (LIST_EMPTY(&lc->lro_free))
return (TCP_LRO_NO_ENTRIES);
/* Start a new segment chain. */
le = LIST_FIRST(&lc->lro_free);
LIST_REMOVE(le, next);
tcp_lro_active_insert(lc, bucket, le);
getmicrotime(&le->mtime);
/* Start filling in details. */
switch (eh_type) {
#ifdef INET6
case ETHERTYPE_IPV6:
le->le_ip6 = ip6;
le->source_ip6 = ip6->ip6_src;
le->dest_ip6 = ip6->ip6_dst;
le->eh_type = eh_type;
le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
break;
#endif
#ifdef INET
case ETHERTYPE_IP:
le->le_ip4 = ip4;
le->source_ip4 = ip4->ip_src.s_addr;
le->dest_ip4 = ip4->ip_dst.s_addr;
le->eh_type = eh_type;
le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
break;
#endif
2008-06-11 22:12:50 +00:00
}
le->source_port = th->th_sport;
le->dest_port = th->th_dport;
le->next_seq = seq + tcp_data_len;
le->ack_seq = th->th_ack;
le->window = th->th_win;
if (l != 0) {
le->timestamp = 1;
le->tsval = ntohl(*(ts_ptr + 1));
le->tsecr = *(ts_ptr + 2);
}
#ifdef TCP_LRO_UPDATE_CSUM
/*
* Do not touch the csum of the first packet. However save the
* "adjusted" checksum of just the source and destination addresses,
* the next header and the TCP payload. The length and TCP header
* parts may change, so we remove those from the saved checksum and
* re-add with final values on tcp_lro_flush() if needed.
*/
KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
__func__, le, le->ulp_csum));
le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
~csum);
th->th_sum = csum; /* Restore checksum on first packet. */
#endif
le->m_head = m;
le->m_tail = m_last(m);
return (0);
2008-06-11 22:12:50 +00:00
}
int
tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
{
return tcp_lro_rx2(lc, m, csum, 1);
}
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
void
tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
{
/* sanity checks */
if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
lc->lro_mbuf_max == 0)) {
/* packet drop */
m_freem(mb);
return;
}
/* check if packet is not LRO capable */
if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
(lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
/* input packet to network layer */
(*lc->ifp->if_input) (lc->ifp, mb);
return;
}
/* create sequence number */
lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
(((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
(((uint64_t)mb->m_pkthdr.flowid) << 24) |
((uint64_t)lc->lro_mbuf_count);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
/* enter mbuf */
lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
/* flush if array is full */
if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
tcp_lro_flush_all(lc);
Add optimizing LRO wrapper: - Add optimizing LRO wrapper which pre-sorts all incoming packets according to the hash type and flowid. This prevents exhaustion of the LRO entries due to too many connections at the same time. Testing using a larger number of higher bandwidth TCP connections showed that the incoming ACK packet aggregation rate increased from ~1.3:1 to almost 3:1. Another test showed that for a number of TCP connections greater than 16 per hardware receive ring, where 8 TCP connections was the LRO active entry limit, there was a significant improvement in throughput due to being able to fully aggregate more than 8 TCP stream. For very few very high bandwidth TCP streams, the optimizing LRO wrapper will add CPU usage instead of reducing CPU usage. This is expected. Network drivers which want to use the optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of "tcp_lro_flush()". Further the LRO control structure must be initialized using "tcp_lro_init_args()" passing a non-zero number into the "lro_mbufs" argument. - Make LRO statistics 64-bit. Previously 32-bit integers were used for statistics which can be prone to wrap-around. Fix this while at it and update all SYSCTL's which expose LRO statistics. - Ensure all data is freed when destroying a LRO control structures, especially leftover LRO entries. - Reduce number of memory allocations needed when setting up a LRO control structure by precomputing the total amount of memory needed. - Add own memory allocation counter for LRO. - Bump the FreeBSD version to force recompilation of all KLDs due to change of the LRO control structure size. Sponsored by: Mellanox Technologies Reviewed by: gallatin, sbruno, rrs, gnn, transport Tested by: Netflix Differential Revision: https://reviews.freebsd.org/D4914
2016-01-19 15:33:28 +00:00
}
/* end */